Here’s a really nice example of high-throughput reaction discovery/condition scouting from a team at Merck. They certainly state the problem correctly:
Modern organic synthesis and especially transition metal catalysis is redefining the rules with which new bonds can be forged, however, it is important to recognize that many “solved” synthetic transformations are far from universal, performing well on simple model substrates yet often failing when applied to complex substrates in real-world synthesis. A recent analysis of 2149 metal catalyzed C-N couplings run in the final steps of the synthesis of highly functionalized drug leads reveals that 55% of reactions failed to deliver any product at all. The missed opportunity represented by these unsuccessful syntheses haunts contemporary drug discovery, and there is a growing recognition that the tendency of polar, highly functionalized compounds to fail in catalysis may actually enrich compound sets in greasy molecules that are less likely to become successful drug candidates. . .
That “recent analysis” they mention turns out to be an internal study of Merck’s own electronic lab notebooks, and it sounds very believable. That’s a problem of organic chemistry: we can do a lot, but rarely can we do it in a general fashion. The paper details an effort to look for Pd-catalyzed coupling reactions in DMSO or NMP, which are (as the authors point out) not the usual solvents that people choose. But they has many advantages for high-throughput experimentation, not least the solubility of more complicated substrates. They started off by screening bases and catalysts in glass microvials in a 96-well array, but then tried those conditions (and more) in a 1536-well plastic plate.
On that level of miniaturization, you can really start clearing some brush. And they uncovered a range of reaction conditions that have not been reported before, using a very real-world set of coupling partners (shown). Applying one of the more general-looking protocols to the whole set, though, still showed about a 50% failure rate, so they turned around and took 32 of the failures and ran new arrays with them of 48 reaction conditions each. (That’s what I mean by clearing things out quickly!) Those 48 reactions consume less than 1 mg of substate in total. By careful mass-based encoding of the array, they could analyze the 1536-well plate in under three hours by LC/MS.
That led to optimized conditions for 21 of the 32, but they took 6 of the remaining recalcitrant combinations and tried another array on them, this time varying catalyst loading, amount of nucleophile, and amount of base. 5 of the 6 yielded to that optimization, which confirms the usual belief that just about any metal-catalyzed coupling will work, if you’re just willing to devote enough of your life to optimizing it. And this automated system significantly changes the value of “enough of your life”.
This is different from Design-of-Experiments setups, in that those are modeled in a way to minimize the number of experimental runs by identifying (or trying to identify) the key variables. But with very small, highly automated experiments, that’s not really as big a concern. You can just let it rip; try a bunch of stuff and look for granularity in the reaction condition space that you’d miss by trying to get more efficient. The Merck team winds up by saying “In biomedical research, chemical synthesis should not limit access to any molecule that is designed to answer a biological question”, and that really is the ideal we should be working towards.